How Christine McDannell built a voice agent and left it switched off

The most useful data point in voice AI right now is a decision not to deploy. Christine McDannell, founder of The Magnolia Firm, an M&A brokerage in Austin, cut payroll 83%, built her own email and meeting agents, pays for two frontier subscriptions, and built a voice agent for the firm that she has not turned on. Her reason is latency — the pause between a caller finishing and the agent answering.

"We didn't fully install it yet, because I'm like: people are gonna just get mad. So let's wait until it's better."

She said it to Artificial Reality, the AI buyer-evidence podcast hosted by Galina Fendikevich. She has not disclosed which vendor built the system.

Is the technology good enough?

By her own account it has crossed the plausibility line. A year and a half before the interview she took a call after joining a gym and could not tell what she was speaking to:

"It was a year and a half ago where I got a gym membership ... And then I got a call. It's like: you just joined, would you like to book your first class? And I'm just having a conversation. Yeah, okay. And then I was like — I think this is a phone agent, like an AI agent. Because it was so hard for me to tell. And then I asked, and it said it was. I was like: whoa, I'm so impressed."

Her assessment of her own agent's voice was positive too — "she sounds friendly, I like that" — and she noted the system could add typing sounds in the background, a feature a colleague turned off. What stops it is the gap:

"The latency though still is — versus awkward long pauses. Like those are annoying."

In a demonstration played during the interview, the agent introduced itself and handled a request about a listing with a noticeable gap before responding to a straightforward question.

Why is a pause worse than a wrong answer?

Because in voice, interaction quality gates the value entirely. A text tool that takes four seconds is fine. Four seconds of silence on a phone call reads as a broken system or an inattentive person, and the caller's judgment forms before any content arrives. That is why partial readiness in this category means unusable rather than useful-with-caveats, and why a firm that automated most of its back office still leaves this one switched off.

McDannell has seen the problem solved. At a hackathon, a teammate built a voice system on a newer model:

"It was zero latency, and it switched the language ... it was speaking English, but then we pretended we were speaking Spanish and it switched to Spanish. It was nuts, and it's zero latency."

She had not used that model herself at the time of the interview: "I haven't played with that model yet."

What would make her deploy it?

Her firm's use case is narrow — an inbound caller is either a buyer or a seller, and routing them correctly with the right information is most of the value. What she wants before switching it on:

• No pause. Latency is the single blocking issue, and she has seen zero-latency systems exist.
• A natural, friendly voice, which she credits her current system with already.
• Correct qualification across a two-sided caller base, where buyer and seller need entirely different follow-up questions.
• Disclosure at the start of the call that the caller is speaking to AI.
• An immediate route to a human, on a keypress.
• A quality bar that keeps callers from getting angry, which is the threshold she named explicitly.

Should the caller be told it is AI?

Both people in the interview said yes. Fendikevich proposed telling callers up front and asking for patience, so the pauses read as expected rather than broken. She is blunt about the current experience generally: "It's not good, it's creepy, especially for some of the sales AI avatars. It just doesn't sound natural, and you end up wanting to be like: operator, operator, get me to a human."

McDannell went further:

"I agree a hundred and ten percent. So I would probably put the disclosure right at the beginning. Hey, I'm the Magnolia Firm's AI assistant and I'm gonna help you. But if at any moment you can push zero to get a person, I would do that."

She notes opinion is divided on disclosure, and offers a reason people currently tolerate voice agents at all: "I think it's such a novelty still right now that people probably just play along."

What a builder should take from this

That novelty tolerance is a depreciating asset, and every mediocre agent shipped today spends it — not only the deployer's, but the category's. By the time the technology is genuinely ready, callers may have learned that an AI voice means a frustrating call, and the good implementations will be paying for the bad ones.

Which makes McDannell's wait look less like caution and more like a commercial read. If you sell voice agents, stop benchmarking against demos and start publishing the number she is actually waiting on: how many milliseconds of silence does your agent leave after the caller stops talking?

Where this comes from

S1E5: How she cut her payroll by 83% using one AI agent — the full interview with Christine McDannell. Listen or watch: YouTube, Spotify or Apple Podcasts.